bool OclPostProcessTransform::prepareKernels()
{
m_kernelFlipH = prepareKernel("transform_flip_h");
m_kernelFlipV = prepareKernel("transform_flip_v");
m_kernelRot180 = prepareKernel("transform_rot_180");
m_kernelRot90 = prepareKernel("transform_rot_90");
m_kernelRot270 = prepareKernel("transform_rot_270");
m_kernelFlipHRot90 = prepareKernel("transform_flip_h_rot_90");
m_kernelFlipVRot90 = prepareKernel("transform_flip_v_rot_90");
return (m_kernelFlipH != NULL)
&& (m_kernelFlipV != NULL)
&& (m_kernelRot180 != NULL)
&& (m_kernelRot90 != NULL)
&& (m_kernelRot270 != NULL)
&& (m_kernelFlipHRot90 != NULL)
&& (m_kernelFlipVRot90 != NULL);
}
void Network::step()
{
if (!built) build();
if (built) {
LayerMap::iterator it;
Layer* l;
//update the external layers
for (it = layers.begin(); it != layers.end(); it++) {
l = it->second;
if (l->isExternal()) {
//update layer and add outputs to native outputs array
ExternalLayer* el = (ExternalLayer*) l;
el->update();
UnitMap::iterator uit;
UnitId unitId;
for (uit = el->units.begin(); uit != el->units.end(); uit++) {
output[unitId] = el->getOutput(unitId);
}
}
}
//run the kernel
cl_int err;
size_t workSize = (size_t) numNonExUnits;
prepareKernel();
err = clEnqueueNDRangeKernel(clDevice.commandQueue, *clKernel, 1, NULL, &workSize, NULL, 0, NULL, NULL);
err = clFinish(clDevice.commandQueue);
if (err != CL_SUCCESS) {
printf("Failed to execute kernel!\n");
}
//copy the new output to the previous output
clEnqueueCopyBuffer(clDevice.commandQueue, clNewOutput, clPrevOutput, 0, 0, sizeof(float)*units.size(), 0, NULL, NULL);
err = clFinish(clDevice.commandQueue);
}
}